Immunefi CEO declares crypto's "vulnerability apocalypse" as AI tilts the cyber battlefield 🛡️

A surge in sophisticated artificial intelligence models has triggered a "vulnerability apocalypse" that is reshaping cybersecurity in favor of attackers, according to Mitchell Amador, CEO of bug bounty platform Immunefi. Speaking with Cointelegraph at the WAIB Summit in Monaco, Amador pointed to the proliferation of models including Claude Opus 4.8 and ChatGPT 5.5 as the main reason behind a resurgence in crypto hacks observed in 2026.

The scale of the threat was underscored by April 2026 data from DefiLlama, which recorded more than $634 million stolen from cryptocurrency platforms that month. That figure marked the highest monthly total since the Bybit hack helped drive industry losses to roughly $1.4 billion in February 2025.

Amador said the next three to four years will be a critical survival period for the crypto sector, until defensive teams can harness the same AI capabilities to build "impregnable" codebases that attackers cannot breach. He added that timeline could shrink to under two years if the industry more widely adopted "crowdsourced security solutions" while researchers work to turn the models to their advantage.

His remarks followed the release of Anthropic's latest Claude Mythos model, Fable 5, which prompted industry concerns over its potential ability to accelerate cryptocurrency exploits. Anthropic said on Tuesday that Fable 5 includes safeguards that reroute topics such as cybersecurity to a different model, Claude Opus 4.8.

The wider industry has grown increasingly alert to protocol risks after a series of major DeFi exploits. On April 19, an attacker drained about 116,500 restaked Ether (rsETH), worth roughly $290 million to $293 million at the time, from Kelp DAO's LayerZero-powered rsETH bridge. LayerZero attributed the breach to Kelp DAO's 1/1 decentralized verifier network (DVN) setup, which created a single point of failure by relying on one verifier path for cross-chain messages, and said it had previously advised against that configuration.